Presentations

Abstract

Reliable wind power forecasting beyond the medium-range (15 days) remains a significant challenge, particularly in mid-latitude regions where weather variability is high. However, estimating the tendency of wind farm production several months in advance is crucial for ensuring stable electricity supply to consumers, especially during winter when demand peaks and wind resources are critical. This study presents a machine learning approach for seasonal wind energy forecasting at timescales of 1 to 6 months ahead. We use seasonal forecast data from ECMWF as input to a gradient boosting model targeting monthly energy production from several wind farms in Pays de la Loire region of France. The model employs quantile regression to predict the full distribution of possible outcomes, enabling explicit quantification of forecast uncertainty. Given the importance of avoiding underproduction for grid stability, we emphasize predictions of lower quantiles, which provide more operationally relevant information than mean or median forecasts alone. The model is validated over a 9-year period from 2017 to 2025. We assess how the predicted probability of exceedance indicators of monthly wind farm production P50 and P90 compare with the climatological references. The one-month ahead P90 values, representing production levels exceeded 90% of the time, are particularly valuable for grid stability and management as it helps prepare for potential underproduction scenarios. The overall skill of the forecast is evaluated using the Continuous Ranked Probability Score (CRPS), which measures both accuracy and sharpness of probabilistic predictions. Calibration is assessed through reliability diagrams comparing forecast quantiles against observed occurrences of monthly productions below the forecast value. Results demonstrate that skillful and reliable seasonal energy forecasts are achievable: at 1-month lead time, we obtain an average CRPS improvement of 30% over the climatological baseline, with skill remaining substantial (27% improvement) even at 6-month lead time.